v0.1.0-RC

Release Notes

This is a release candidate for the CINN project.

Major features

• APIs
  • Added NetBuilder and Compile APIs to help build a PaddlePadle model and compile and execute it
  • Added CinnBuilder to build a tensor computation based on some fine-grained operations
  • Added CINN DSL to make it easy to write any tensor computation
• Codegen and backend
  • Support for execution on CPU and Nvidia GPU with two modes
    1. JIT mode based on LLVM backend and NVRTC
    2. source code mode with generating C and CUDA source code
• Scheduling
  • Similar to Halide and TVM projects, introduced the following schedule directives:
    • fuse, split
    • reorder, tile
    • unroll, compute_inline
    • bind, compute_at
    • parallel, vectorize
    • cache_read, cache_write
• Training support
  • Embedded in PaddlePaddle training framework and supported Deep Learning training
• Inference support
  • Support for loading a PaddlePaddle model and executing and getting the result
• Install and build
  • Added C++ library for C++ stack usage
  • Added Python whl package for Python stack usage

Document

• Added several documents, please take a look at CINN docs

Also thanks to our external contributors

Thanks to our external contributors @fc500110 @T8T9

v0.1-alpha

v0.1-alpha

This is a pre-release version and provides an alpha version of the project CINN.